Add transitional learning detect with in-dialog mask review.

Senses grow from confirmations instead of a word list, so a later local pipeline can decide homonyms on a short passage. Default stays semantic. Local detect waits longer, and the reverse-proxy timeout is documented so Dev does not 504 first.

Co-authored-by: Cursor <cursoragent@cursor.com>

backend/detect_learning.py

@@ -0,0 +1,410 @@
+"""Transitional learning detect: senses grow from dialogue confirmations.
+
+Not a global word list. Detect still does not auto-activate identities.
+Ambiguous spellings get a local passage call when a local detect provider exists.
+"""
+from __future__ import annotations
+
+import json
+import re
+import uuid
+from typing import Any
+
+from db import get_db, row_to_dict
+from identity_store import ENTITY_TYPES, confirm_identity, is_maskable_label, normalize_label
+
+MODE_SEMANTIC = "semantic"
+MODE_LEARNING = "learning"
+MODES = (MODE_SEMANTIC, MODE_LEARNING)
+SETTING_KEY = "detect_operating_mode"
+_LETTER = r"A-Za-zÄÖÜäöüß"
+WINDOW = 80
+
+_DECISION_IDENTITY = "identity"
+_DECISION_NOT = "not_identity"
+
+
+def get_detect_operating_mode() -> str:
+    with get_db() as conn:
+        row = row_to_dict(
+            conn.execute("SELECT value FROM app_settings WHERE key = ?", (SETTING_KEY,)).fetchone()
+        )
+    value = ((row or {}).get("value") or MODE_SEMANTIC).strip().lower()
+    return value if value in MODES else MODE_SEMANTIC
+
+
+def set_detect_operating_mode(mode: str) -> str:
+    chosen = (mode or "").strip().lower()
+    if chosen not in MODES:
+        raise ValueError("invalid_detect_operating_mode")
+    with get_db() as conn:
+        conn.execute(
+            """
+            INSERT INTO app_settings (key, value, updated)
+            VALUES (?, ?, datetime('now'))
+            ON CONFLICT(key) DO UPDATE SET value = excluded.value, updated = datetime('now')
+            """,
+            (SETTING_KEY, chosen),
+        )
+    return chosen
+
+
+def get_sense(profile_id: str, label: str) -> dict:
+    key = normalize_label(label) or (label or "").strip()
+    if not key:
+        return {
+            "normalized_label": "",
+            "identity_hits": 0,
+            "non_identity_hits": 0,
+            "ambiguous": False,
+        }
+    with get_db() as conn:
+        row = row_to_dict(
+            conn.execute(
+                """
+                SELECT normalized_label, identity_hits, non_identity_hits
+                FROM label_senses
+                WHERE profile_id = ? AND lower(normalized_label) = lower(?)
+                """,
+                (profile_id, key),
+            ).fetchone()
+        )
+    if not row:
+        return {
+            "normalized_label": key,
+            "identity_hits": 0,
+            "non_identity_hits": 0,
+            "ambiguous": False,
+        }
+    identity = int(row.get("identity_hits") or 0)
+    other = int(row.get("non_identity_hits") or 0)
+    return {
+        "normalized_label": row.get("normalized_label") or key,
+        "identity_hits": identity,
+        "non_identity_hits": other,
+        "ambiguous": identity > 0 and other > 0,
+    }
+
+
+def record_sense(profile_id: str, label: str, *, identity: bool) -> dict:
+    key = normalize_label(label) or (label or "").strip()
+    if not key:
+        raise ValueError("empty_label")
+    with get_db() as conn:
+        existing = row_to_dict(
+            conn.execute(
+                """
+                SELECT normalized_label, identity_hits, non_identity_hits FROM label_senses
+                WHERE profile_id = ? AND lower(normalized_label) = lower(?)
+                """,
+                (profile_id, key),
+            ).fetchone()
+        )
+        if existing:
+            key = existing.get("normalized_label") or key
+        identity_hits = int((existing or {}).get("identity_hits") or 0)
+        non_identity_hits = int((existing or {}).get("non_identity_hits") or 0)
+        if identity:
+            identity_hits += 1
+        else:
+            non_identity_hits += 1
+        conn.execute(
+            """
+            INSERT INTO label_senses (
+              profile_id, normalized_label, identity_hits, non_identity_hits, updated
+            )
+            VALUES (?, ?, ?, ?, datetime('now'))
+            ON CONFLICT(profile_id, normalized_label) DO UPDATE SET
+              identity_hits = excluded.identity_hits,
+              non_identity_hits = excluded.non_identity_hits,
+              updated = datetime('now')
+            """,
+            (profile_id, key, identity_hits, non_identity_hits),
+        )
+    return get_sense(profile_id, key)
+
+
+def _in_user_text(label: str, user_body: str) -> bool:
+    if not label or not user_body:
+        return False
+    return bool(
+        re.search(
+            rf"(?<![{_LETTER}]){re.escape(label)}(?![{_LETTER}])",
+            user_body,
+            re.IGNORECASE,
+        )
+    )
+
+
+def _excerpt(user_body: str, start: int | None, end: int | None, label: str) -> str:
+    text = user_body or ""
+    if start is None or end is None or start < 0 or end > len(text):
+        match = re.search(rf"(?<![{_LETTER}]){re.escape(label)}(?![{_LETTER}])", text, re.IGNORECASE)
+        if not match:
+            return text[:WINDOW]
+        start, end = match.start(), match.end()
+    left = max(0, int(start) - WINDOW)
+    right = min(len(text), int(end) + WINDOW)
+    return text[left:right]
+
+
+def needs_review(profile_id: str, mapping: dict, user_body: str) -> bool:
+    label = mapping.get("local_label") or ""
+    if mapping.get("source") == "confirmed_registry":
+        sense = get_sense(profile_id, label)
+        return bool(sense["ambiguous"]) and _in_user_text(label, user_body)
+    if mapping.get("source") != "request_local":
+        return False
+    if not is_maskable_label(label) or not _in_user_text(label, user_body):
+        return False
+    sense = get_sense(profile_id, label)
+    if sense["identity_hits"] > 0 and not sense["ambiguous"]:
+        return False
+    return True
+
+
+def build_candidates(profile_id: str, mappings: list[dict], user_body: str) -> list[dict]:
+    seen: set[tuple[int | None, int | None, str]] = set()
+    items: list[dict] = []
+    for mapping in mappings:
+        if not needs_review(profile_id, mapping, user_body):
+            continue
+        label = mapping.get("local_label") or ""
+        start = mapping.get("start")
+        end = mapping.get("end")
+        key = (start, end, label.casefold())
+        if key in seen:
+            continue
+        seen.add(key)
+        items.append(
+            {
+                "id": str(uuid.uuid4()),
+                "text": label,
+                "entity_type": mapping.get("entity_type") or "PERSON",
+                "start": start,
+                "end": end,
+                "excerpt": _excerpt(user_body, start, end, label),
+                "ambiguous": get_sense(profile_id, label)["ambiguous"],
+            }
+        )
+    return items
+
+
+def suppress_known_non_identity(profile_id: str, mappings: list[dict], user_body: str) -> list[dict]:
+    kept = []
+    for mapping in mappings:
+        label = mapping.get("local_label") or ""
+        sense = get_sense(profile_id, label)
+        if (
+            mapping.get("source") == "request_local"
+            and sense["non_identity_hits"] > 0
+            and not sense["identity_hits"]
+            and _in_user_text(label, user_body)
+        ):
+            continue
+        kept.append(mapping)
+    return kept
+
+
+def save_pending(
+    profile_id: str,
+    conversation_id: str,
+    user_message_id: str,
+    payload: dict,
+) -> str:
+    review_id = str(uuid.uuid4())
+    with get_db() as conn:
+        conn.execute(
+            "DELETE FROM pending_mask_reviews WHERE profile_id = ? AND conversation_id = ?",
+            (profile_id, conversation_id),
+        )
+        conn.execute(
+            """
+            INSERT INTO pending_mask_reviews (
+              id, profile_id, conversation_id, user_message_id, payload_json, created
+            )
+            VALUES (?, ?, ?, ?, ?, datetime('now'))
+            """,
+            (review_id, profile_id, conversation_id, user_message_id, json.dumps(payload, ensure_ascii=False)),
+        )
+    return review_id
+
+
+def load_pending(profile_id: str, review_id: str) -> dict | None:
+    with get_db() as conn:
+        row = row_to_dict(
+            conn.execute(
+                """
+                SELECT * FROM pending_mask_reviews
+                WHERE id = ? AND profile_id = ?
+                """,
+                (review_id, profile_id),
+            ).fetchone()
+        )
+    if not row:
+        return None
+    payload = json.loads(row.get("payload_json") or "{}")
+    payload["id"] = row["id"]
+    payload["conversation_id"] = row["conversation_id"]
+    payload["user_message_id"] = row["user_message_id"]
+    return payload
+
+
+def drop_pending(profile_id: str, review_id: str) -> None:
+    with get_db() as conn:
+        conn.execute(
+            "DELETE FROM pending_mask_reviews WHERE id = ? AND profile_id = ?",
+            (review_id, profile_id),
+        )
+
+
+def apply_review_decisions(profile_id: str, pending: dict, decisions: list[dict]) -> list[dict]:
+    by_id = {item["id"]: item for item in pending.get("candidates") or []}
+    drop_keys: set[tuple] = set()
+    for raw in decisions:
+        candidate = by_id.get(raw.get("id") or "")
+        if not candidate:
+            continue
+        decision = (raw.get("decision") or "").strip()
+        label = candidate.get("text") or ""
+        kind = candidate.get("entity_type") if candidate.get("entity_type") in ENTITY_TYPES else "PERSON"
+        if decision == _DECISION_IDENTITY:
+            record_sense(profile_id, label, identity=True)
+            confirm_identity(profile_id, label, entity_type=kind)
+        elif decision == _DECISION_NOT:
+            record_sense(profile_id, label, identity=False)
+            drop_keys.add((candidate.get("start"), candidate.get("end"), label.casefold()))
+    mappings = []
+    for mapping in pending.get("mappings") or []:
+        key = (mapping.get("start"), mapping.get("end"), (mapping.get("local_label") or "").casefold())
+        if key in drop_keys:
+            continue
+        mappings.append(mapping)
+    return suppress_known_non_identity(profile_id, mappings, pending.get("user_body") or "")
+
+
+def confirm_known_identity_spans(profile_id: str, mappings: list[dict], user_body: str) -> None:
+    for mapping in mappings:
+        label = mapping.get("local_label") or ""
+        sense = get_sense(profile_id, label)
+        if not (
+            sense["identity_hits"] > 0
+            and not sense["ambiguous"]
+            and _in_user_text(label, user_body)
+        ):
+            continue
+        kind = mapping.get("entity_type") if mapping.get("entity_type") in ENTITY_TYPES else "PERSON"
+        try:
+            confirm_identity(profile_id, label, entity_type=kind)
+        except ValueError:
+            continue
+
+
+def list_senses(profile_id: str) -> list[dict]:
+    with get_db() as conn:
+        rows = conn.execute(
+            """
+            SELECT normalized_label, identity_hits, non_identity_hits
+            FROM label_senses WHERE profile_id = ?
+            ORDER BY lower(normalized_label)
+            """,
+            (profile_id,),
+        ).fetchall()
+    items = []
+    for raw in rows:
+        row = row_to_dict(raw) or {}
+        identity = int(row.get("identity_hits") or 0)
+        other = int(row.get("non_identity_hits") or 0)
+        items.append(
+            {
+                "normalized_label": row.get("normalized_label") or "",
+                "identity_hits": identity,
+                "non_identity_hits": other,
+                "ambiguous": identity > 0 and other > 0,
+            }
+        )
+    return items
+
+
+def pending_for_conversation(profile_id: str, conversation_id: str) -> dict | None:
+    with get_db() as conn:
+        row = row_to_dict(
+            conn.execute(
+                """
+                SELECT id, payload_json FROM pending_mask_reviews
+                WHERE profile_id = ? AND conversation_id = ?
+                """,
+                (profile_id, conversation_id),
+            ).fetchone()
+        )
+    if not row:
+        return None
+    payload = json.loads(row.get("payload_json") or "{}")
+    return {
+        "id": row["id"],
+        "candidates": payload.get("candidates") or [],
+    }
+
+
+def auto_resolve_ambiguous(profile_id: str, candidates: list[dict], mappings: list[dict]) -> tuple[list[dict], list[dict]]:
+    """Local passage LLM for already-ambiguous spellings. Unresolved stay for the popup."""
+    remaining = []
+    drop_keys: set[tuple] = set()
+    for item in candidates:
+        if not item.get("ambiguous"):
+            remaining.append(item)
+            continue
+        label = item.get("text") or ""
+        decision = try_local_passage_decision(item.get("excerpt") or "", label)
+        if decision == _DECISION_NOT:
+            record_sense(profile_id, label, identity=False)
+            drop_keys.add((item.get("start"), item.get("end"), label.casefold()))
+        elif decision == _DECISION_IDENTITY:
+            record_sense(profile_id, label, identity=True)
+        else:
+            remaining.append(item)
+    if not drop_keys:
+        return remaining, mappings
+    kept = []
+    for mapping in mappings:
+        key = (mapping.get("start"), mapping.get("end"), (mapping.get("local_label") or "").casefold())
+        if key in drop_keys:
+            continue
+        kept.append(mapping)
+    return remaining, kept
+
+
+def try_local_passage_decision(excerpt: str, label: str) -> str | None:
+    """Return identity, not_identity, or None if no local model or unusable answer."""
+    from providers import ProviderError, complete_chat, detect_provider
+
+    config = detect_provider()
+    if not config or not config.local or config.mode != "http":
+        return None
+    prompt = (
+        "Entscheide nur für die markierte Nennung in diesem kurzen Ausschnitt. "
+        "Ist sie eine schützenswerte Identität (Person, Ort, Organisation, privates Projekt) "
+        f"oder eine Sache/Allgemeinbedeutung? Wort: {label}\n"
+        f"Ausschnitt: {excerpt}\n"
+        'Antworte nur mit JSON {"decision":"identity"} oder {"decision":"not_identity"}.'
+    )
+    try:
+        result = complete_chat(
+            config,
+            [{"role": "user", "content": prompt}],
+            timeout=30.0,
+            max_tokens=32,
+        )
+    except ProviderError:
+        return None
+    match = re.search(r"\{.*\}", result.content or "", re.DOTALL)
+    if not match:
+        return None
+    try:
+        data = json.loads(match.group(0))
+    except json.JSONDecodeError:
+        return None
+    value = (data.get("decision") or "").strip()
+    if value in {_DECISION_IDENTITY, _DECISION_NOT}:
+        return value
+    return None

backend/dialogue_turn.py

@@ -7,9 +7,21 @@ import re
 from context_builder import assemble_text, build_internal_context, is_closing_turn
 from conversation_signals import infer_signals
 from debug_store import persist_engine_error, persist_step
+from detect_learning import (
+    MODE_LEARNING,
+    apply_review_decisions,
+    auto_resolve_ambiguous,
+    build_candidates,
+    confirm_known_identity_spans,
+    drop_pending,
+    get_detect_operating_mode,
+    load_pending,
+    save_pending,
+    suppress_known_non_identity,
+)
 from dialogue_store import append_message, get_conversation, list_messages, update_conversation_signals
-from engine import EngineError, execute_prompt, load_active_prompt
-from entity_detect import DETECT_DIALOGUE_FALLBACK_CODES
+from engine import EngineError, execute_prompt, load_active_prompt, preview_prompt
+from entity_detect import DETECT_DIALOGUE_FALLBACK_CODES, DetectError, detect_personal_egress
 from writing_profile_store import remember_dialogue_style
 from profile_review import consider_dialogue
 
@@ -302,23 +314,21 @@ def visible_for_role(payload: dict, role: str | None) -> dict:
     return cleaned
 
 
-def run_turn(profile_id: str, conversation_id: str, body: str, message_id: str | None = None) -> dict:
-    conversation = get_conversation(profile_id, conversation_id)
-    user = append_message(profile_id, conversation_id, body, role="user", message_id=message_id)
-    context = build_internal_context(
-        profile_id,
-        conversation_id=conversation_id,
-        space_id=conversation.get("space_id"),
-        journal_day_id=conversation.get("journal_day_id"),
-        purpose="dialogue_turn",
-    )
-    assembled = assemble_text(context)
-    prompt = load_active_prompt("mvp.dialogue_turn")
+def _finish_turn(
+    profile_id: str,
+    conversation_id: str,
+    user: dict,
+    assembled: dict,
+    prompt: dict,
+    *,
+    precomputed_mappings: list[dict] | None = None,
+) -> dict:
     calls = 0
     result = None
     impulse = ""
     decision: dict = {"operation": "unparsed", "label": "nicht erkannt", "parsed": False}
     call_traces: list[dict] = []
+    working = dict(assembled)
     try:
         while calls < 2:
             result = execute_prompt(
@@ -326,7 +336,8 @@ def run_turn(profile_id: str, conversation_id: str, body: str, message_id: str |
                 profile_id,
                 purpose="dialogue_turn",
                 data_class="B",
-                context=assembled,
+                context=working,
+                precomputed_mappings=precomputed_mappings if calls == 0 else None,
             )
             calls += 1
             if result.get("trace"):
@@ -335,16 +346,16 @@ def run_turn(profile_id: str, conversation_id: str, body: str, message_id: str |
             if not content:
                 raise EngineError("empty_provider_response", "Der Provider lieferte keine Antwort.")
             impulse, decision = parse_turn_payload(content)
-            if not needs_repair(impulse, assembled):
+            if not needs_repair(impulse, working):
                 break
             decision = {**decision, "guard": "impulse_rejected"}
-            assembled = dict(assembled)
-            assembled["dialogue_context"] = (
-                (assembled.get("dialogue_context") or "") + "\n\n" + repair_note(impulse, assembled)
+            working = dict(working)
+            working["dialogue_context"] = (
+                (working.get("dialogue_context") or "") + "\n\n" + repair_note(impulse, working)
             )
-        if needs_repair(impulse, assembled):
+        if needs_repair(impulse, working):
             parsed_ok = bool(decision.get("parsed"))
-            impulse = local_hold(last_user_text(assembled))
+            impulse = local_hold(last_user_text(working))
             decision = {
                 "operation": "fortfuehren",
                 "label": OPERATIONS["fortfuehren"],
@@ -369,7 +380,7 @@ def run_turn(profile_id: str, conversation_id: str, body: str, message_id: str |
                 },
             )
             raise
-        impulse = local_hold(last_user_text(assembled))
+        impulse = local_hold(last_user_text(working))
         decision = {
             "operation": "fortfuehren",
             "label": OPERATIONS["fortfuehren"],
@@ -388,7 +399,7 @@ def run_turn(profile_id: str, conversation_id: str, body: str, message_id: str |
         infer_signals(user_bodies, decision.get("operation")),
     )
     remember_dialogue_style(profile_id)
-    consider_dialogue(profile_id, body)
+    consider_dialogue(profile_id, user.get("body") or "")
     trace = result.get("trace") if result else None
     persist_step(
         profile_id,
@@ -416,3 +427,99 @@ def run_turn(profile_id: str, conversation_id: str, body: str, message_id: str |
         "decision": decision,
         "trace": trace,
     }
+
+
+def _learning_pause(profile_id: str, conversation_id: str, user: dict, assembled: dict, prompt: dict) -> dict | None:
+    preview = preview_prompt(prompt, assembled)
+    rendered = preview.get("rendered") or ""
+    try:
+        outcome = detect_personal_egress(profile_id, rendered)
+    except DetectError:
+        return None
+    user_body = user.get("body") or ""
+    mappings = list(outcome.mappings or [])
+    candidates = build_candidates(profile_id, mappings, user_body)
+    candidates, mappings = auto_resolve_ambiguous(profile_id, candidates, mappings)
+    if not candidates:
+        confirm_known_identity_spans(profile_id, mappings, user_body)
+        mappings = suppress_known_non_identity(profile_id, mappings, user_body)
+        return _finish_turn(
+            profile_id,
+            conversation_id,
+            user,
+            assembled,
+            prompt,
+            precomputed_mappings=mappings,
+        )
+    review_id = save_pending(
+        profile_id,
+        conversation_id,
+        user.get("id") or "",
+        {
+            "mappings": mappings,
+            "candidates": candidates,
+            "user_body": user_body,
+        },
+    )
+    return {
+        "conversation": get_conversation(profile_id, conversation_id),
+        "user": user,
+        "assistant": None,
+        "calls": 0,
+        "messages": list_messages(profile_id, conversation_id),
+        "pending_mask_review": {
+            "id": review_id,
+            "candidates": candidates,
+        },
+    }
+
+
+def run_turn(profile_id: str, conversation_id: str, body: str, message_id: str | None = None) -> dict:
+    conversation = get_conversation(profile_id, conversation_id)
+    user = append_message(profile_id, conversation_id, body, role="user", message_id=message_id)
+    context = build_internal_context(
+        profile_id,
+        conversation_id=conversation_id,
+        space_id=conversation.get("space_id"),
+        journal_day_id=conversation.get("journal_day_id"),
+        purpose="dialogue_turn",
+    )
+    assembled = assemble_text(context)
+    prompt = load_active_prompt("mvp.dialogue_turn")
+    if get_detect_operating_mode() == MODE_LEARNING:
+        paused = _learning_pause(profile_id, conversation_id, user, assembled, prompt)
+        if paused is not None:
+            return paused
+    return _finish_turn(profile_id, conversation_id, user, assembled, prompt)
+
+
+def continue_turn(profile_id: str, conversation_id: str, review_id: str, decisions: list[dict]) -> dict:
+    pending = load_pending(profile_id, review_id)
+    if not pending or pending.get("conversation_id") != conversation_id:
+        raise EngineError("mask_review_missing", "Die Maskierungsprüfung ist nicht mehr gültig.", 404)
+    mappings = apply_review_decisions(profile_id, pending, decisions)
+    drop_pending(profile_id, review_id)
+    conversation = get_conversation(profile_id, conversation_id)
+    context = build_internal_context(
+        profile_id,
+        conversation_id=conversation_id,
+        space_id=conversation.get("space_id"),
+        journal_day_id=conversation.get("journal_day_id"),
+        purpose="dialogue_turn",
+    )
+    assembled = assemble_text(context)
+    prompt = load_active_prompt("mvp.dialogue_turn")
+    user = {"id": pending.get("user_message_id"), "body": pending.get("user_body") or ""}
+    messages = list_messages(profile_id, conversation_id)
+    for item in messages:
+        if item.get("id") == pending.get("user_message_id"):
+            user = item
+            break
+    return _finish_turn(
+        profile_id,
+        conversation_id,
+        user,
+        assembled,
+        prompt,
+        precomputed_mappings=mappings,
+    )

backend/engine.py

@@ -75,6 +75,7 @@ def execute_prompt(
     disable_context_compression: bool = False,
     budget=None,
     diagnostics: dict[str, Any] | None = None,
+    precomputed_mappings: list[dict] | None = None,
 ) -> dict:
     preview = preview_prompt(prompt, context)
     feature_id = prompt.get("required_feature") or "ai_calls"
@@ -107,6 +108,7 @@ def execute_prompt(
                     "disable_context_compression": disable_context_compression,
                     "budget": budget,
                     "diagnostics": diagnostics or {},
+                    "precomputed_mappings": precomputed_mappings,
                 },
             )
         )

backend/entity_detect.py

@@ -31,7 +31,9 @@ from providers import ChatResult, ProviderError, complete_chat, detect_provider
 DETECT_CHUNK_CHARS = 4000
 DETECT_CHUNK_OVERLAP = 250
 DETECT_TIMEOUT = 90.0
+DETECT_TIMEOUT_LOCAL = 300.0
 DETECT_MAX_TOKENS = 1024
+DETECT_MAX_TOKENS_LOCAL = 256
 JSON_BLOCK = re.compile(r"\{.*\}", re.DOTALL)
 ALLOWED_ENTITY_FIELDS = frozenset({"start", "end", "text", "entity_type"})
 ALLOWED_ROOT_FIELDS = frozenset({"entities"})
@@ -573,6 +575,13 @@ def _add_usage(stats: DetectionStats, usage: dict | None) -> None:
             pass
 
 
+def detect_chat_limits(config) -> tuple[float, int]:
+    """Remote detect stays at 90s/1024. Local CPU hosts need a longer wait and a shorter cap."""
+    if getattr(config, "local", False):
+        return DETECT_TIMEOUT_LOCAL, DETECT_MAX_TOKENS_LOCAL
+    return DETECT_TIMEOUT, DETECT_MAX_TOKENS
+
+
 def _llm_chunk(config, excerpt: str, *, schema_retry: bool = False) -> ChatResult:
     prompt = resolve_template(
         _detect_prompt()["template"],
@@ -580,11 +589,12 @@ def _llm_chunk(config, excerpt: str, *, schema_retry: bool = False) -> ChatResul
     )
     if schema_retry:
         prompt = f"{prompt.rstrip()}\n\n{SCHEMA_RETRY_HINT}"
+    timeout, max_tokens = detect_chat_limits(config)
     return complete_chat(
         config,
         [{"role": "user", "content": prompt}],
-        timeout=DETECT_TIMEOUT,
-        max_tokens=DETECT_MAX_TOKENS,
+        timeout=timeout,
+        max_tokens=max_tokens,
         disable_context_compression=True,
     )
 

backend/migrations/024_detect_learning.sql

@@ -0,0 +1,23 @@
+-- Transitional learning detect: user-confirmed label senses, pending mask review.
+-- Detect output still does not auto-activate identities.
+
+CREATE TABLE IF NOT EXISTS label_senses (
+  profile_id TEXT NOT NULL REFERENCES profiles(id) ON DELETE CASCADE,
+  normalized_label TEXT NOT NULL,
+  identity_hits INTEGER NOT NULL DEFAULT 0,
+  non_identity_hits INTEGER NOT NULL DEFAULT 0,
+  updated TEXT NOT NULL DEFAULT CURRENT_TIMESTAMP::text,
+  PRIMARY KEY (profile_id, normalized_label)
+);
+
+CREATE TABLE IF NOT EXISTS pending_mask_reviews (
+  id TEXT PRIMARY KEY,
+  profile_id TEXT NOT NULL REFERENCES profiles(id) ON DELETE CASCADE,
+  conversation_id TEXT NOT NULL REFERENCES conversations(id) ON DELETE CASCADE,
+  user_message_id TEXT NOT NULL,
+  payload_json TEXT NOT NULL DEFAULT '{}',
+  created TEXT NOT NULL DEFAULT CURRENT_TIMESTAMP::text
+);
+
+CREATE INDEX IF NOT EXISTS idx_pending_mask_reviews_conv
+  ON pending_mask_reviews (profile_id, conversation_id);

backend/privacy_gateway.py

@@ -1055,7 +1055,26 @@ def complete(request: GatewayRequest) -> GatewayResult:
     detect_note = None
     detect_stats: dict[str, Any] = {}
     local_identities: list[dict[str, Any]] = []
+    precomputed = request.payload.get("precomputed_mappings") if request.payload else None
     try:
+        if precomputed is not None:
+            mappings = list(precomputed)
+            detect_stats = {
+                "detect_note": "precomputed_learning_review",
+                "full_detection_coverage": True,
+                "semantic_identity_guaranteed": False,
+            }
+            local_identities = [
+                {
+                    "local_label": item.get("local_label"),
+                    "token": item.get("token"),
+                    "entity_type": item.get("entity_type"),
+                    "source": item.get("source"),
+                }
+                for item in mappings
+            ]
+            detect_name = "learning_review"
+        else:
             outcome = detect_personal_egress(request.profile_id, rendered)
             mappings = outcome.mappings
             detect_stats = outcome.stats.public()

backend/provider_settings.py

@@ -8,6 +8,7 @@ from pathlib import Path
 from urllib.parse import urlparse
 
 from db import get_db
+from detect_learning import get_detect_operating_mode
 from env_loader import (
     allows_remote_plaintext_detect,
     remote_plaintext_detect_reason,
@@ -460,4 +461,5 @@ def public_status() -> dict:
         "remote_plaintext_reason": remote_plaintext_detect_reason(),
         "remote_plaintext_allowed": allows_remote_plaintext_detect(),
         "ollama_url": OLLAMA_LAN_URL,
+        "detect_operating_mode": get_detect_operating_mode(),
     }

backend/routers/admin.py

@@ -20,6 +20,7 @@ from debug_store import (
     settings_payload,
 )
 from dialogue_store import StoreError, get_conversation, inventory, list_conversations, list_derived_for_conversation, list_messages
+from detect_learning import list_senses
 from identity_store import (
     confirm_identity,
     confirm_review_proposal,
@@ -226,6 +227,28 @@ class ActivateProfile(BaseModel):
     profile_id: str
 
 
+class DetectModeWrite(BaseModel):
+    mode: str
+
+
+@router.put("/providers/detect-mode")
+def admin_set_detect_mode(body: DetectModeWrite, session: dict = Depends(require_admin_dep)):
+    from detect_learning import set_detect_operating_mode
+    from provider_settings import public_status as provider_public_status
+
+    try:
+        set_detect_operating_mode(body.mode)
+    except ValueError:
+        raise HTTPException(
+            status_code=400,
+            detail={
+                "code": "invalid_detect_operating_mode",
+                "message": "Detect-Modus muss semantic oder learning sein.",
+            },
+        )
+    return provider_public_status()
+
+
 @router.get("/providers")
 def admin_providers(session: dict = Depends(require_admin_dep)):
     return public_status()
@@ -364,6 +387,7 @@ def admin_identities(session: dict = Depends(require_admin_dep)):
     return {
         "registry": list_registry(profile_id),
         "proposals": list_review_proposals(profile_id, include_dismissed=True),
+        "senses": list_senses(profile_id),
         "note": "Nur lokale bestätigte Registry und unbestätigte Vorschläge. Kein externer Egress.",
         "backup": (
             "Vor einer Bereinigung die lokale Datei backend/data/kansho.sqlite kopieren. "

backend/routers/dialogue.py

@@ -7,7 +7,8 @@ from auth import require_auth
 from context_builder import build_internal_context
 from continuity import checkpoint_usage_session, close_usage_session
 from derived_kinds import catalog
-from dialogue_turn import run_turn, visible_for_role
+from detect_learning import pending_for_conversation
+from dialogue_turn import continue_turn, run_turn, visible_for_role
 from engine import EngineError
 from privacy_gateway import public_error_detail
 from dialogue_store import (
@@ -51,6 +52,11 @@ class MessageWrite(BaseModel):
     id: str | None = None
 
 
+class MaskReviewWrite(BaseModel):
+    review_id: str
+    decisions: list[dict] = Field(default_factory=list)
+
+
 class ThreadWrite(BaseModel):
     title: str = ""
     status: str = "open"
@@ -140,6 +146,9 @@ def get_one_conversation(conversation_id: str, session: dict = Depends(require_a
         conv = get_conversation(session["profile_id"], conversation_id)
         conv["messages"] = list_messages(session["profile_id"], conversation_id)
         conv["derived"] = list_derived_for_conversation(session["profile_id"], conversation_id)
+        pending = pending_for_conversation(session["profile_id"], conversation_id)
+        if pending:
+            conv["pending_mask_review"] = pending
         return conv
     except StoreError as exc:
         _http(exc)
@@ -172,6 +181,17 @@ def conversation_turn(conversation_id: str, req: MessageWrite, session: dict = D
         _http(exc)
 
 
+@router.post("/conversations/{conversation_id}/turn/review")
+def conversation_turn_review(conversation_id: str, req: MaskReviewWrite, session: dict = Depends(require_auth)):
+    try:
+        return visible_for_role(
+            continue_turn(session["profile_id"], conversation_id, req.review_id, req.decisions),
+            session.get("role"),
+        )
+    except (StoreError, EngineError) as exc:
+        _http(exc)
+
+
 @router.get("/conversations/{conversation_id}/messages")
 def get_messages(conversation_id: str, session: dict = Depends(require_auth)):
     try:

backend/routers/journal.py

@@ -5,8 +5,9 @@ from fastapi.responses import FileResponse
 from pydantic import BaseModel, Field
 
 from auth import require_auth
+from detect_learning import pending_for_conversation
 from dialogue_store import StoreError, delete_conversation, get_conversation, list_messages
-from dialogue_turn import run_turn, visible_for_role
+from dialogue_turn import continue_turn, run_turn, visible_for_role
 from privacy_gateway import GatewayRequest, inspect, public_error_detail
 from engine import EngineError
 from journal_generate import generate_draft
@@ -96,6 +97,11 @@ class TurnWrite(BaseModel):
     id: str | None = None
 
 
+class MaskReviewWrite(BaseModel):
+    review_id: str
+    decisions: list[dict] = Field(default_factory=list)
+
+
 class GenerateWrite(BaseModel):
     conversation_ids: list[str] | None = None
     include_existing: bool = False
@@ -286,7 +292,11 @@ def remove_conversation(conversation_id: str, session: dict = Depends(require_au
 def read_conversation(conversation_id: str, session: dict = Depends(require_auth)):
     try:
         conv = get_conversation(session["profile_id"], conversation_id)
-        return {"conversation": conv, "messages": list_messages(session["profile_id"], conversation_id)}
+        payload = {"conversation": conv, "messages": list_messages(session["profile_id"], conversation_id)}
+        pending = pending_for_conversation(session["profile_id"], conversation_id)
+        if pending:
+            payload["pending_mask_review"] = pending
+        return payload
     except StoreError as exc:
         _http(exc)
 
@@ -308,6 +318,17 @@ def conversation_turn(conversation_id: str, body: TurnWrite, session: dict = Dep
         _http(exc)
 
 
+@router.post("/conversations/{conversation_id}/turn/review")
+def conversation_turn_review(conversation_id: str, body: MaskReviewWrite, session: dict = Depends(require_auth)):
+    try:
+        return visible_for_role(
+            continue_turn(session["profile_id"], conversation_id, body.review_id, body.decisions),
+            session.get("role"),
+        )
+    except (StoreError, EngineError) as exc:
+        _http(exc)
+
+
 @router.get("/generation-settings")
 def read_generation_settings(session: dict = Depends(require_auth)):
     return settings_payload(session["profile_id"])

backend/schema.sql

@@ -230,6 +230,27 @@ CREATE TABLE IF NOT EXISTS identity_review_proposals (
   UNIQUE (profile_id, observed_label, entity_type)
 );
 
+CREATE TABLE IF NOT EXISTS label_senses (
+  profile_id TEXT NOT NULL REFERENCES profiles(id) ON DELETE CASCADE,
+  normalized_label TEXT NOT NULL,
+  identity_hits INTEGER NOT NULL DEFAULT 0,
+  non_identity_hits INTEGER NOT NULL DEFAULT 0,
+  updated TEXT NOT NULL DEFAULT (datetime('now')),
+  PRIMARY KEY (profile_id, normalized_label)
+);
+
+CREATE TABLE IF NOT EXISTS pending_mask_reviews (
+  id TEXT PRIMARY KEY,
+  profile_id TEXT NOT NULL REFERENCES profiles(id) ON DELETE CASCADE,
+  conversation_id TEXT NOT NULL REFERENCES conversations(id) ON DELETE CASCADE,
+  user_message_id TEXT NOT NULL,
+  payload_json TEXT NOT NULL DEFAULT '{}',
+  created TEXT NOT NULL DEFAULT (datetime('now'))
+);
+
+CREATE INDEX IF NOT EXISTS idx_pending_mask_reviews_conv
+  ON pending_mask_reviews (profile_id, conversation_id);
+
 CREATE TABLE IF NOT EXISTS journal_days (
   id TEXT PRIMARY KEY,
   profile_id TEXT NOT NULL REFERENCES profiles(id) ON DELETE CASCADE,

backend/sqlite_to_postgres.py

@@ -43,6 +43,8 @@ TABLES = [
     "re_grounding_events",
     "identity_mappings",
     "identity_review_proposals",
+    "label_senses",
+    "pending_mask_reviews",
     "journal_days",
     "journal_drafts",
     "journal_entries",

backend/tests/test_detect_learning.py

@@ -0,0 +1,184 @@
+"""Transitional learning detect: senses from dialogue review, admin mode switch.
+
+Fake detect/provider only. Run from backend/: python tests/test_detect_learning.py
+"""
+from __future__ import annotations
+
+import os
+import sys
+from pathlib import Path
+from unittest.mock import patch
+
+ROOT = Path(__file__).resolve().parents[1]
+sys.path.insert(0, str(ROOT))
+
+os.environ["KANSHO_FAKE_PROVIDER"] = "1"
+os.environ["KANSHO_FAKE_DETECT"] = "1"
+
+from fastapi.testclient import TestClient
+from detect_learning import get_detect_operating_mode, get_sense, record_sense
+from dialogue_turn import visible_for_role
+from identity_store import list_confirmed_identities
+from main import app
+from privacy_gateway import reset_debug
+
+
+def expect(ok: bool, message: str) -> None:
+    if not ok:
+        raise SystemExit(f"FAIL: {message}")
+    print(f"OK  {message}")
+
+
+def status_ok(response, label: str) -> None:
+    if response.status_code != 200:
+        detail = response.json() if response.headers.get("content-type", "").startswith("application/json") else {}
+        code = detail.get("detail", {}).get("code") if isinstance(detail, dict) else None
+        expect(False, f"{label} HTTP {response.status_code} {code or ''}".strip())
+    expect(True, label)
+
+
+def header(token: str) -> dict:
+    return {"X-Auth-Token": token}
+
+
+def open_conv(client: TestClient, headers: dict) -> str:
+    space = client.post("/api/journal/spaces", headers=headers, json={"title": "Alltag"})
+    day = client.post(
+        f"/api/journal/spaces/{space.json()['id']}/days",
+        headers=headers,
+        json={"calendar_date": "2026-09-08"},
+    )
+    conv = client.post(
+        f"/api/journal/days/{day.json()['day']['id']}/conversations",
+        headers=headers,
+        json={"title": "Gespräch"},
+    )
+    return conv.json()["id"]
+
+
+def main() -> None:
+    reset_debug()
+    with TestClient(app) as client:
+        setup = client.post(
+            "/api/auth/setup",
+            json={"email": "lars@example.test", "name": "Lars", "password": "test-pass"},
+        )
+        headers = header(setup.json()["token"])
+        profile_id = setup.json()["profile_id"]
+
+        providers = client.get("/api/admin/providers", headers=headers)
+        expect(providers.status_code == 200, "providers status")
+        expect(providers.json()["detect_operating_mode"] == "semantic", "default detect mode is semantic")
+        expect(get_detect_operating_mode() == "semantic", "store default is semantic")
+
+        bad = client.put("/api/admin/providers/detect-mode", headers=headers, json={"mode": "wordlist"})
+        expect(bad.status_code == 400, "invalid detect mode is rejected")
+
+        conv_semantic = open_conv(client, headers)
+        semantic_turn = client.post(
+            f"/api/journal/conversations/{conv_semantic}/turn",
+            headers=headers,
+            json={"body": "Anna kam vorbei."},
+        )
+        status_ok(semantic_turn, "semantic turn")
+        expect(not semantic_turn.json().get("pending_mask_review"), "semantic mode does not pause for review")
+        expect(any(item.get("role") == "assistant" for item in semantic_turn.json().get("messages") or []), "semantic turn gets an assistant")
+
+        switched = client.put("/api/admin/providers/detect-mode", headers=headers, json={"mode": "learning"})
+        status_ok(switched, "switch to learning")
+        expect(switched.json()["detect_operating_mode"] == "learning", "learning mode persisted")
+
+        conv = open_conv(client, headers)
+        paused = client.post(
+            f"/api/journal/conversations/{conv}/turn",
+            headers=headers,
+            json={"body": "Anna kam vorbei."},
+        )
+        status_ok(paused, "learning pause")
+        review = paused.json().get("pending_mask_review") or {}
+        expect(bool(review.get("id")), "learning mode returns a review id")
+        expect(paused.json().get("assistant") is None, "paused turn has no assistant")
+        candidates = review.get("candidates") or []
+        anna = next((item for item in candidates if item.get("text") == "Anna"), None)
+        expect(anna is not None, "Anna is a review candidate")
+        hidden = visible_for_role({"pending_mask_review": review, "trace": {"egress": "x"}}, "user")
+        expect("pending_mask_review" in hidden, "users still receive the mask review")
+        expect("trace" not in hidden, "users do not receive the detect trace")
+
+        reloaded = client.get(f"/api/journal/conversations/{conv}", headers=headers)
+        expect(reloaded.json().get("pending_mask_review", {}).get("id") == review["id"], "pending review survives reload")
+
+        missing = client.post(
+            f"/api/journal/conversations/{conv}/turn/review",
+            headers=headers,
+            json={"review_id": "missing", "decisions": []},
+        )
+        expect(missing.status_code == 404, "unknown review is gone")
+
+        identity = client.post(
+            f"/api/journal/conversations/{conv}/turn/review",
+            headers=headers,
+            json={
+                "review_id": review["id"],
+                "decisions": [{"id": anna["id"], "decision": "identity"}],
+            },
+        )
+        status_ok(identity, "identity review")
+        expect(not identity.json().get("pending_mask_review"), "review continue clears the pause")
+        expect(any(item.get("role") == "assistant" for item in identity.json().get("messages") or []), "continue produces an assistant")
+        confirmed = list_confirmed_identities(profile_id)
+        expect(any((item.get("canonical_label") or "").casefold() == "anna" for item in confirmed), "identity review confirms the registry")
+        sense = get_sense(profile_id, "Anna")
+        expect(sense["identity_hits"] >= 1 and not sense["ambiguous"], "Anna is identity-only")
+
+        second = client.post(
+            f"/api/journal/conversations/{conv}/turn",
+            headers=headers,
+            json={"body": "Anna hat später angerufen."},
+        )
+        status_ok(second, "second Anna turn")
+        expect(not second.json().get("pending_mask_review"), "identity-only spelling skips the popup")
+
+        food_conv = open_conv(client, headers)
+        food = client.post(
+            f"/api/journal/conversations/{food_conv}/turn",
+            headers=headers,
+            json={"body": "Hanna aß mit uns."},
+        )
+        status_ok(food, "hanna pause")
+        food_review = food.json().get("pending_mask_review") or {}
+        food_candidates = food_review.get("candidates") or []
+        hanna = next((item for item in food_candidates if item.get("text") == "Hanna"), None)
+        expect(hanna is not None, "Hanna is a review candidate")
+        declined = client.post(
+            f"/api/journal/conversations/{food_conv}/turn/review",
+            headers=headers,
+            json={
+                "review_id": food_review["id"],
+                "decisions": [{"id": hanna["id"], "decision": "not_identity"}],
+            },
+        )
+        status_ok(declined, "not_identity review")
+        expect(not any((item.get("canonical_label") or "").casefold() == "hanna" for item in list_confirmed_identities(profile_id)), "not_identity does not confirm Hanna")
+        expect(get_sense(profile_id, "Hanna")["non_identity_hits"] >= 1, "not_identity records a sense")
+
+        record_sense(profile_id, "Clarissa", identity=True)
+        record_sense(profile_id, "Clarissa", identity=False)
+        expect(get_sense(profile_id, "Clarissa")["ambiguous"], "both senses mark Clarissa as ambiguous")
+        amb_conv = open_conv(client, headers)
+        with patch("detect_learning.try_local_passage_decision", return_value="not_identity"):
+            amb = client.post(
+                f"/api/journal/conversations/{amb_conv}/turn",
+                headers=headers,
+                json={"body": "Clarissa lag auf dem Tisch."},
+            )
+        status_ok(amb, "ambiguous local passage")
+        expect(not amb.json().get("pending_mask_review"), "local passage can finish without a popup")
+        expect(any(item.get("role") == "assistant" for item in amb.json().get("messages") or []), "local passage still generates")
+
+        back = client.put("/api/admin/providers/detect-mode", headers=headers, json={"mode": "semantic"})
+        expect(back.json()["detect_operating_mode"] == "semantic", "mode can return to semantic")
+
+
+if __name__ == "__main__":
+    main()

backend/tests/test_frame.py

@@ -89,6 +89,7 @@ def main() -> None:
         expect(set(roles) == {"generate", "detect"}, "two provider roles")
         expect(providers.json()["remote_plaintext_allowed"] is True, "test runtime allows remote detect")
         expect(providers.json()["remote_plaintext_reason"] == "non_production", "tests are not production")
+        expect(providers.json().get("detect_operating_mode") == "semantic", "detect mode defaults to semantic")
         expect(roles["generate"]["ready"] is False, "generate fail-closed without key")
         expect("sk-" not in providers.text, "provider status has no secret")
         expect("KANSHO_PROVIDER_KEY=" not in providers.text, "env assignment not leaked")

backend/tests/test_local_url.py

@@ -12,7 +12,14 @@ os.environ.setdefault("KANSHO_PROVIDER_KEY", "")
 os.environ.setdefault("KANSHO_DETECT_PROVIDER_KEY", "")
 Path(tempfile.gettempdir()).mkdir(parents=True, exist_ok=True)
 
-from providers import is_local_url
+from entity_detect import (
+    DETECT_MAX_TOKENS,
+    DETECT_MAX_TOKENS_LOCAL,
+    DETECT_TIMEOUT,
+    DETECT_TIMEOUT_LOCAL,
+    detect_chat_limits,
+)
+from providers import ProviderConfig, is_local_url
 
 
 def expect(ok: bool, message: str) -> None:
@@ -29,6 +36,33 @@ def main() -> None:
     expect(is_local_url("http://ollama:11434/v1/chat/completions"), "compose hostname ollama is local")
     expect(not is_local_url("https://openrouter.ai/api/v1/chat/completions"), "OpenRouter is not local")
     expect(not is_local_url("https://8.8.8.8/v1/chat/completions"), "public IP is not local")
+    remote = ProviderConfig(
+        role="detect",
+        name="openrouter",
+        mode="http",
+        url="https://openrouter.ai/api/v1/chat/completions",
+        model="openai/gpt-4.1-nano",
+        key="x",
+        local=False,
+        zdr=True,
+        no_train=True,
+    )
+    local = ProviderConfig(
+        role="detect",
+        name="ollama",
+        mode="http",
+        url="http://192.168.2.144:11434/v1/chat/completions",
+        model="phi3:mini",
+        key="",
+        local=True,
+        zdr=True,
+        no_train=True,
+    )
+    expect(detect_chat_limits(remote) == (DETECT_TIMEOUT, DETECT_MAX_TOKENS), "remote detect keeps 90s/1024")
+    expect(
+        detect_chat_limits(local) == (DETECT_TIMEOUT_LOCAL, DETECT_MAX_TOKENS_LOCAL),
+        "local detect waits 300s with a 256-token cap",
+    )
     print("local url tests passed.")
 
 

docs/DEPLOYMENT.md

@@ -49,6 +49,8 @@ cp .env.example .env
 
 Host-nginx: `nginx/kansho.conf` und `nginx/kansho-dev.conf` nach `/etc/nginx/sites-available/`, dann `nginx/certbot-setup.sh`. DNS A/AAAA für `dev.kansho.jinkendo.de` und `kansho.jinkendo.de` auf den Reverse-Proxy. Bis TLS steht, bleibt der LAN-Zugriff über die Publish-Ports.
 
+**Additiv 2026-09-08:** Live-TLS endet auf dem Synology-NAS `192.168.2.63` (DSM Reverse Proxy, nicht auf dem Kanshō-Pi). Pro Kanshō-Regel unter Erweitert Proxy-Lese- und -Sende-Timeout auf 600s. DSM-Default 60s liefert HTTP 504, während Detect/Generate auf dem Pi weiterlaufen. LAN `http://192.168.2.49:3096` umgeht diesen Hop (Frontend-Container bereits 600s). AdGuard-Rewrites bleiben auf `192.168.2.63`.
+
 Gitea: Actions aktivieren. Derselbe Pi-Runner wie die Schwesterprodukte (`ubuntu-latest`).
 
 Watchtower bleibt aus.

docs/architecture/functional/guardrails.md

@@ -709,6 +709,18 @@ Das ist kein Admin-Feature-Flag und kein Umdeuten von `KANSHO_ENV=production` na
 
 **Später prüfen:** Flag entfernen, sobald lokales Detect auf Prod läuft. Dann wieder Production ohne Klartext-Detect.
 
+## 22.4 Übergang: Detect-Lernmodus (2026-09-08)
+
+Additiv. Technische Abbildung: `../technical/privacy_gateway.md` §9.7.
+
+**Entschieden (Übergang):** Doppeldeutigkeiten werden nicht als eigene Wortliste gepflegt. Sinne entstehen aus Bestätigungen im Dialog (Identität / nicht schützenswert). Beides bei derselben Schreibweise markiert sie als mehrdeutig.
+
+**Entschieden (Übergang):** Im Admin konfigurierbarer Detector-Modus `semantic` (Default) oder `learning`. Lernmodus öffnet vor Generate ein Bestätigungs-Popup für Detect-Treffer im aktuellen Nutzersatz. Das ist mehr Arbeit am Anfang und kein Ersatz für semantische Detection.
+
+**Entschieden (Übergang):** Eine Mini-Passage an ein internes Modell geht nur bei bereits mehrdeutiger Schreibweise und nur an ein lokales Detect. OpenRouter sieht diese Passage nicht. Fehlt das lokale Modell, bleibt das Popup.
+
+**Nicht:** Gateway abschalten, Detect-Treffer auto-speichern, `Sushi_`/`Sushi+` im Nutzertext, Pattern-Wortliste als Wahrheit.
+
 ---
 
 # 23. Externe Referenzquellen

docs/architecture/technical/admin_diagnostics.md

@@ -55,6 +55,8 @@ Nicht übernehmen: Mitai-Admin für Körpertarife, Coupons, Training Types als K
 
 **Additiv 2026-09-08 (LLM-Profile):** `GET/POST/PUT/DELETE /api/admin/llm-profiles`, `POST /api/admin/providers/{role}/activate`. Profile enthalten URL/Modell/ZDR, niemals Keys. LAN-Ollama (`192.168.2.144:11434`) ist lokal.
 
+**Additiv 2026-09-08 (Detect-Lernmodus):** `PUT /api/admin/providers/detect-mode` mit `semantic` | `learning`. Statusfeld `detect_operating_mode`. Keine Wortlisten-Seite. Identitäten dürfen ein Badge „mehrdeutig“ aus Dialog-Sinnen zeigen. Dialog und Journal-Gespräch pausieren im Lernmodus mit Bestätigungs-Popup; Journal-Generate nicht.
+
 ## 4.2 Implementierungsstand (Dialog-Testspur)
 
 **Additiv 2026-08-26:** Lokale Identitätsregistry unter `/admin/identities`. Detect-Vorschläge sind unbestätigt. Compact-Diagnose enthält Detect-Abdeckung, Chunks, Kosten und Laufzeit, aber keine Labels.

docs/architecture/technical/mvp_implementation.md

@@ -179,7 +179,7 @@ Editor speichert Markdown, nicht HTML. Medien-Token bleiben lokal. Dirty: In-App
 
 | Schicht | Rolle | Call |
 |---|---|---|
-| Maskierung | `detect` | Vollständige semantische Detection des Generate-Egress. Externes Klartext-Detect in Development/Test; in Production default-off, optional Operator-Übergang `KANSHO_ALLOW_REMOTE_DETECT`. Ziel: lokales Detect-Modell. Kein Pattern-Fallback. |
+| Maskierung | `detect` | Vollständige semantische Detection des Generate-Egress. Externes Klartext-Detect in Development/Test; in Production default-off, optional Operator-Übergang `KANSHO_ALLOW_REMOTE_DETECT`. Ziel: lokales Detect-Modell. Kein Pattern-Fallback. Admin-Modus `semantic` (Default) oder `learning` (Übergang, Bestätigung vor Generate). |
 | Dialogzug | `generate` | Operation + Impuls |
 | Journalentwurf | `generate` | Explizit, getrennt |
 

docs/architecture/technical/privacy_gateway.md

@@ -239,7 +239,7 @@ Bestätigte Registry-Zeilen und bestätigte Aliase werden im gesamten gerenderte
 
 ### Tests und Live-Qualität
 
-Contract-Tests: `backend/tests/test_privacy_detect.py`, `backend/tests/test_identity_registry.py`, `backend/tests/test_privacy_response_integrity.py`, `backend/tests/test_detect_contract_retry.py`. Sie beweisen Schema, Fail-closed, span-genaue Maskierung und Datenfluss, nicht semantische Modellleistung. Opt-in: `python entity_detect_eval.py --live` mit synthetischen Sätzen und exakten erwarteten Spans. Ohne diesen Lauf bleibt die Live-Qualität unbestätigt. Das aktuell konfigurierte `openai/gpt-4.1-nano` gilt durch reale False-Positive-Vorschläge qualitativ nicht als zuverlässig bestätigt; das Modell wird deshalb nicht stillschweigend gewechselt.
+Contract-Tests: `backend/tests/test_privacy_detect.py`, `backend/tests/test_identity_registry.py`, `backend/tests/test_privacy_response_integrity.py`, `backend/tests/test_detect_contract_retry.py`, `backend/tests/test_detect_learning.py`. Sie beweisen Schema, Fail-closed, span-genaue Maskierung, Lernmodus-Pause und Datenfluss, nicht semantische Modellleistung. Opt-in: `python entity_detect_eval.py --live` mit synthetischen Sätzen und exakten erwarteten Spans. Ohne diesen Lauf bleibt die Live-Qualität unbestätigt. Das aktuell konfigurierte `openai/gpt-4.1-nano` gilt durch reale False-Positive-Vorschläge qualitativ nicht als zuverlässig bestätigt; das Modell wird deshalb nicht stillschweigend gewechselt.
 
 ## 9.6 Übergang: `KANSHO_ALLOW_REMOTE_DETECT` (2026-09-08)
 
@@ -251,7 +251,21 @@ Operator-Übergang bis Ollama: `KANSHO_ALLOW_REMOTE_DETECT=1` (Compose `.env`, C
 
 **Additiv 2026-09-08 (LLM-Profile, LAN-Ollama):** `llm_profiles` speichert benannte URL/Modell-Presets ohne Keys. Jede Stufe (`generate` / `detect`) zeigt auf ein Profil; Wechsel kopiert die gespeicherten Felder, statt sie neu einzugeben. RFC1918-, Loopback- und `.local`-URLs gelten als lokale Trusted Zone. `http://192.168.2.144:11434/v1/chat/completions` ist damit lokales Detect, kein externes Klartext-Detect. Ollama muss auf dem Host auf `0.0.0.0:11434` lauschen; der Pi muss Port 11434 erreichen.
 
-Status in `GET /api/admin/providers`: `remote_plaintext_allowed`, `remote_plaintext_reason` (`non_production` | `operator_override` | `blocked`), `profiles`. Modellauswahl: `GET /api/admin/providers/models` (OpenRouter-Katalog oder Ollama `/api/tags`, soft-fail).
+**Additiv 2026-09-08 (lokales Detect, Timeouts):** CPU-Ollama braucht für denselben Detect-Prompt deutlich länger als OpenRouter. Remote-Detect bleibt 90s / 1024 Tokens. Lokales Detect wartet 300s und begrenzt auf 256 Tokens. Live-TLS endet auf dem Synology-Reverse-Proxy (`192.168.2.63`, DSM-Default 60s). Dort Proxy-Lese- und -Sende-Timeout 600s setzen, sonst sieht der Browser HTTP 504, während das Backend weiterläuft. Ein Dialogzug kann danach lokal einen Halte-Impuls gespeichert haben. Die Compose-Frontend-Nginx auf Port 3096 hat bereits 600s.
+
+Status in `GET /api/admin/providers`: `remote_plaintext_allowed`, `remote_plaintext_reason` (`non_production` | `operator_override` | `blocked`), `profiles`, `detect_operating_mode` (`semantic` | `learning`). Modellauswahl: `GET /api/admin/providers/models` (OpenRouter-Katalog oder Ollama `/api/tags`, soft-fail).
+
+## 9.7 Übergang: Detect-Lernmodus (2026-09-08)
+
+Additiv. Fachliches Home: `../functional/guardrails.md` §22.4. Ersetzt weder semantische Detection noch die bestätigte Registry.
+
+Default bleibt `semantic`: Detect läuft wie bisher, Generate folgt ohne Pause. `learning` ist ein Admin-Schalter (`PUT /api/admin/providers/detect-mode`), kein Gateway-Bypass und keine Wortlisten-UI.
+
+Im Lernmodus untersucht Detect weiterhin den vollen gerenderten Generate-Egress. Bevor Generate startet, werden `request_local`-Spans im aktuellen Nutzersatz zur Bestätigung angeboten (Dialog und Journal-Gespräch, nicht Journal-Generate). „Identität“ bestätigt die lokale Registry und zählt einen Identitätssinn. „Nicht schützenswert“ maskiert diese Nennung nicht und zählt den anderen Sinn. Detect-Ausgabe allein speichert weiterhin keine aktive Identität.
+
+Es gibt keine separat zu pflegende Doppeldeutigkeitsliste. Mehrdeutigkeit entsteht, wenn dieselbe Schreibweise beide Sinne hat. Nur dann darf ein **lokales** Detect-Modell eine Mini-Passage (Ausschnitt um die Nennung) entscheiden. Fehlt ein lokales Modell oder ist die Antwort unbrauchbar, bleibt das Popup. OpenRouter erhält diese Passage nicht.
+
+Nach der Bestätigung läuft Generate mit den geprüften Mappings (`precomputed_learning_review`), ohne zweiten Detect-Pass. Compact-Diagnose enthält weiterhin keine Klartextlabels. Tests: `backend/tests/test_detect_learning.py`.
 
 ## 10. Offene Fragen
 
@@ -259,6 +273,8 @@ Status in `GET /api/admin/providers`: `remote_plaintext_allowed`, `remote_plaint
 
 **Additiv 2026-08-28:** Wie sollen neue, vom Detect verfehlte und lokal nicht bestätigte Identitäten vor dem Egress behandelt werden, ohne span-genaue Homonyme oder Allgemeinbegriffe dauerhaft zu maskieren? Optionen ohne Vorentscheidung: fail-closed vor Generate, Review der Detect-Vorschläge vor Egress, oder ein zweites lokales Verfahren. Keine Heuristik auf Verdacht.
 
+**Additiv 2026-09-08:** Der Detect-Lernmodus ist eine Übergangslösung für die Review-Option, nicht der Endzustand. Semantische Detection bleibt Pflicht. Unbekannte Namen sind nicht deshalb harmlos, weil sie nicht auf einer Liste stehen.
+
 ## 11. Querverweise
 
 - Fachlich: `../functional/guardrails.md`

docs/architecture/technical/runtime_and_deploy.md

@@ -85,6 +85,8 @@ merge --no-ff develop in main → push main → deploy-prod.yml
 
 **Additiv 2026-09-07 (Qualitätssystem):** Backend-Runner ist pytest (`backend/pytest.ini`, `backend/tests/conftest.py`). Gitea startet die Suite erst nach erfolgreichem Dev-Deploy (`workflow_run` auf `Deploy Development`), nicht parallel zum Image-Build. Destruktive Tests bleiben auf `kansho_test`. Zusätzlich ein nicht-schreibender Smoke gegen die laufende Dev-API (`kansho_dev`, nur `/api/health`). Öffentliche URLs nach Host-Nginx/TLS: `https://dev.kansho.jinkendo.de` und `https://kansho.jinkendo.de`. LAN-Ports 3096/8096 und 3006/8005 bleiben die Compose-Publish-Ziele.
 
+**Additiv 2026-09-08 (TLS-Hop):** DNS für `*.kansho.jinkendo.de` zeigt auf das Synology-NAS `192.168.2.63`. Der Pi hat kein Host-HTTPS auf 443. DSM Reverse Proxy: Proxy-Lese-/Sende-Timeout 600s; Default 60s bricht lokale Detect-Läufe mit HTTP 504 ab. Ziel bleibt der Pi (`3096`/`8096` Dev, `3006`/`8005` Prod). Repo-Dateien `nginx/kansho.conf` und `nginx/kansho-dev.conf` beschreiben denselben Timeout, falls TLS später auf dem Pi endet.
+
 Kanshō-Repo liegt auf Gitea (`Lars/Kansho`). HTTPS-Push ist eingerichtet. Der Pi-Runner (`ubuntu-latest`) ist derselbe wie bei Mitai/Shinkan/Kairo.
 
 ## 5. Was Deploy nicht übernimmt

frontend/src/app.css

@@ -687,3 +687,11 @@ pre.code {
 .runlog-list li.bad { border-color: #e3c0c0; }
 .runlog-trace { margin-top: 0.8rem; }
 .runlog-trace .trace-panel { border-top: 1px dashed var(--line); }
+.runlog-modal fieldset.row-actions {
+  border: 0;
+  margin: 0.4rem 0 0;
+  padding: 0;
+  display: flex;
+  flex-wrap: wrap;
+  gap: 0.8rem;
+}

frontend/src/components/MaskReviewPanel.jsx

@@ -0,0 +1,53 @@
+export default function MaskReviewPanel({ review, busy, onSubmit }) {
+  if (!review?.candidates?.length) return null
+  return (
+    <div className="runlog-overlay" role="dialog" aria-labelledby="mask-review-title">
+      <div className="runlog-modal">
+        <div className="runlog-head">
+          <div>
+            <h2 id="mask-review-title">Maskierung prüfen</h2>
+            <p className="runlog-status">
+              Lernmodus: nur diese Nennung. Bestätigen legt die Bezeichnung lokal ab.
+              „Nicht schützenswert“ verhindert die Maskierung. Beides bei demselben Wort markiert es als mehrdeutig.
+            </p>
+          </div>
+        </div>
+        <form
+          className="stack"
+          onSubmit={(event) => {
+            event.preventDefault()
+            const data = new FormData(event.currentTarget)
+            const decisions = review.candidates.map((item) => ({
+              id: item.id,
+              decision: data.get(`decision-${item.id}`) || 'identity'
+            }))
+            onSubmit(decisions)
+          }}
+        >
+          <ul className="stack">
+            {review.candidates.map((item) => (
+              <li key={item.id} className="row-item">
+                <div>
+                  <strong>{item.text}</strong>
+                  <span className="muted"> · {item.entity_type}{item.ambiguous ? ' · mehrdeutig' : ''}</span>
+                  {item.excerpt && <p className="muted">{item.excerpt}</p>}
+                </div>
+                <fieldset className="row-actions">
+                  <label className="check">
+                    <input type="radio" name={`decision-${item.id}`} value="identity" defaultChecked />
+                    Identität
+                  </label>
+                  <label className="check">
+                    <input type="radio" name={`decision-${item.id}`} value="not_identity" />
+                    Nicht schützenswert
+                  </label>
+                </fieldset>
+              </li>
+            ))}
+          </ul>
+          <button type="submit" disabled={busy}>{busy ? 'Übernimmt …' : 'Übernehmen und weiter'}</button>
+        </form>
+      </div>
+    </div>
+  )
+}

frontend/src/pages/AdminIdentitiesPage.jsx

@@ -104,6 +104,7 @@ export default function AdminIdentitiesPage() {
       <p className="muted">
         Nur lokale bestätigte Registry. Detect-Treffer werden nicht automatisch aktiv.
         Unbestätigte Vorschläge gelten nicht als bekannte Identität. Kein externer Egress.
+        Mehrdeutigkeit entsteht aus Dialog-Bestätigungen, nicht aus einer Wortlisten-Seite.
       </p>
       {error && <p className="error">{error}</p>}
       {notice && <p>{notice}</p>}
@@ -120,9 +121,16 @@ export default function AdminIdentitiesPage() {
             </tr>
           </thead>
           <tbody>
-            {(data?.registry || []).map((item) => (
+            {(data?.registry || []).map((item) => {
+              const sense = (data?.senses || []).find(
+                (entry) => (entry.normalized_label || '').toLowerCase() === (item.canonical_label || '').toLowerCase()
+              )
+              return (
               <tr key={item.id}>
-                <td>{item.canonical_label}</td>
+                <td>
+                  {item.canonical_label}
+                  {sense?.ambiguous ? <span className="muted"> · mehrdeutig</span> : null}
+                </td>
                 <td>
                   <select
                     value={item.entity_type}
@@ -150,7 +158,8 @@ export default function AdminIdentitiesPage() {
                   <button type="button" className="ghost" onClick={() => remove(item.id)}>Entfernen</button>
                 </td>
               </tr>
-            ))}
+              )
+            })}
           </tbody>
         </table>
       </div>

frontend/src/pages/AdminProvidersPage.jsx

@@ -152,6 +152,27 @@ export default function AdminProvidersPage() {
     }
   }
 
+  const saveDetectMode = async (mode) => {
+    setError('')
+    setNotice('')
+    setSaving('detect-mode')
+    try {
+      const payload = await api('/api/admin/providers/detect-mode', {
+        token: session.token,
+        method: 'PUT',
+        body: { mode }
+      })
+      applyPayload(payload)
+      setNotice(mode === 'learning'
+        ? 'Lernmodus: Detect pausiert vor Generate, bis Maskierungen bestätigt sind.'
+        : 'Detect-Modus: semantisch, ohne Bestätigungs-Popup.')
+    } catch (err) {
+      setError(err.message)
+    } finally {
+      setSaving('')
+    }
+  }
+
   const saveRole = async (role) => {
     const form = forms[role]
     if (!form) return
@@ -255,6 +276,25 @@ export default function AdminProvidersPage() {
               <div key={role.role} className="stack">
                 <h2>{role.title || role.role}</h2>
                 {role.task && <p className="muted">{role.task}</p>}
+                {role.role === 'detect' && (
+                  <label>
+                    Detect-Modus
+                    <select
+                      value={data.detect_operating_mode || 'semantic'}
+                      onChange={(e) => saveDetectMode(e.target.value)}
+                      disabled={saving === 'detect-mode'}
+                    >
+                      <option value="semantic">Semantisch (Default)</option>
+                      <option value="learning">Lernmodus (Übergang)</option>
+                    </select>
+                  </label>
+                )}
+                {role.role === 'detect' && data.detect_operating_mode === 'learning' && (
+                  <p className="muted">
+                    Keine separate Wortliste. Bestätigungen im Dialog legen lokale Sinne an.
+                    Mehrdeutige Nennungen gehen nur als Mini-Passage an ein lokales Detect-Modell.
+                  </p>
+                )}
                 <dl className="meta">
                   <dt>Status</dt>
                   <dd>{statusLabel(view)}</dd>

frontend/src/pages/DialoguePage.jsx

@@ -1,6 +1,7 @@
 import { useEffect, useState } from 'react'
 import { Link } from 'react-router-dom'
 import { api, apiDownload } from '../api.js'
+import MaskReviewPanel from '../components/MaskReviewPanel.jsx'
 import { useAuth } from '../context/AuthContext.jsx'
 
 function convLabel(item, index) {
@@ -18,6 +19,7 @@ export default function DialoguePage() {
   const [busy, setBusy] = useState(false)
   const [egress, setEgress] = useState(null)
   const [debugOn, setDebugOn] = useState(false)
+  const [maskReview, setMaskReview] = useState(null)
 
   const loadList = () =>
     api('/api/dialogue/conversations', { token: session.token })
@@ -33,6 +35,7 @@ export default function DialoguePage() {
       const data = await api(`/api/dialogue/conversations/${id}`, { token: session.token })
       setActiveId(id)
       setMessages(data.messages || [])
+      setMaskReview(data.pending_mask_review || null)
     } catch (err) {
       setError(err.message)
     }
@@ -78,6 +81,7 @@ export default function DialoguePage() {
       else {
         setActiveId(null)
         setMessages([])
+        setMaskReview(null)
       }
     } catch (err) {
       setError(err.message)
@@ -109,7 +113,12 @@ export default function DialoguePage() {
       })
       setDraft('')
       setMessages(result.messages || [])
+      if (result.pending_mask_review) {
+        setMaskReview(result.pending_mask_review)
+      } else {
+        setMaskReview(null)
         await loadList()
+      }
     } catch (err) {
       setError(err.message)
       if (activeId) await openConversation(activeId)
@@ -118,6 +127,26 @@ export default function DialoguePage() {
     }
   }
 
+  const submitMaskReview = async (decisions) => {
+    if (!activeId || !maskReview?.id) return
+    setError('')
+    setBusy(true)
+    try {
+      const result = await api(`/api/dialogue/conversations/${activeId}/turn/review`, {
+        token: session.token,
+        method: 'POST',
+        body: { review_id: maskReview.id, decisions }
+      })
+      setMaskReview(null)
+      setMessages(result.messages || [])
+      await loadList()
+    } catch (err) {
+      setError(err.message)
+    } finally {
+      setBusy(false)
+    }
+  }
+
   return (
     <section className="card dialogue-page">
       <header className="dialogue-head">
@@ -174,11 +203,12 @@ export default function DialoguePage() {
           {activeId && (
             <form className="dialogue-composer" onSubmit={send}>
               <textarea rows={3} value={draft} onChange={(e) => setDraft(e.target.value)} placeholder="Erzählen …" />
-              <button type="submit" disabled={busy}>{busy ? 'Antwortet …' : 'Senden'}</button>
+              <button type="submit" disabled={busy || Boolean(maskReview)}>{busy ? 'Antwortet …' : 'Senden'}</button>
             </form>
           )}
         </div>
       </div>
+      <MaskReviewPanel review={maskReview} busy={busy} onSubmit={submitMaskReview} />
     </section>
   )
 }

frontend/src/pages/JournalDayPage.jsx

@@ -2,6 +2,7 @@ import { useEffect, useState } from 'react'
 import { Link, useNavigate, useParams } from 'react-router-dom'
 import { api, apiDownload } from '../api.js'
 import DayScratch from '../components/DayScratch.jsx'
+import MaskReviewPanel from '../components/MaskReviewPanel.jsx'
 import RunLogPopup from '../components/RunLogPopup.jsx'
 import { useAuth } from '../context/AuthContext.jsx'
 import { entryTitle } from '../journal/document.js'
@@ -58,6 +59,7 @@ export default function JournalDayPage() {
   const [logStatus, setLogStatus] = useState('ok')
   const [runLog, setRunLog] = useState([])
   const [debugOn, setDebugOn] = useState(false)
+  const [maskReview, setMaskReview] = useState(null)
 
   const loadDay = async (preferId) => {
     const data = await api(`/api/journal/days/${dayId}`, { token: session.token })
@@ -68,9 +70,11 @@ export default function JournalDayPage() {
       const conv = await api(`/api/journal/conversations/${nextId}`, { token: session.token })
       setActiveId(nextId)
       setMessages(conv.messages || [])
+      setMaskReview(conv.pending_mask_review || null)
     } else {
       setActiveId(null)
       setMessages([])
+      setMaskReview(null)
     }
     return data
   }
@@ -148,7 +152,12 @@ export default function JournalDayPage() {
       })
       setDraft('')
       setMessages(result.messages || [])
+      if (result.pending_mask_review) {
+        setMaskReview(result.pending_mask_review)
+      } else {
+        setMaskReview(null)
         await loadDay(activeId)
+      }
     } catch (err) {
       setError(err.message)
       await loadDay(activeId)
@@ -157,6 +166,26 @@ export default function JournalDayPage() {
     }
   }
 
+  const submitMaskReview = async (decisions) => {
+    if (!activeId || !maskReview?.id) return
+    setError('')
+    setBusy(true)
+    try {
+      const result = await api(`/api/journal/conversations/${activeId}/turn/review`, {
+        token: session.token,
+        method: 'POST',
+        body: { review_id: maskReview.id, decisions }
+      })
+      setMaskReview(null)
+      setMessages(result.messages || [])
+      await loadDay(activeId)
+    } catch (err) {
+      setError(err.message)
+    } finally {
+      setBusy(false)
+    }
+  }
+
   const generate = async (conversationIds) => {
     setError('')
     setBusy(true)
@@ -391,7 +420,7 @@ export default function JournalDayPage() {
                 onChange={(e) => setDraft(e.target.value)}
                 placeholder="Erzählen …"
               />
-              <button type="submit" disabled={busy}>{busy ? 'Antwortet …' : 'Senden'}</button>
+              <button type="submit" disabled={busy || Boolean(maskReview)}>{busy ? 'Antwortet …' : 'Senden'}</button>
             </form>
           )}
         </div>
@@ -403,6 +432,7 @@ export default function JournalDayPage() {
           onError={setError}
         />
       </div>
+      <MaskReviewPanel review={maskReview} busy={busy} onSubmit={submitMaskReview} />
       <RunLogPopup
         open={logOpen}
         status={logStatus}

nginx/kansho-dev.conf

@@ -1,4 +1,6 @@
 # Kanshō development vhost. Install as /etc/nginx/sites-available/kansho-dev.
+# Live TLS terminates on the Synology Reverse Proxy at 192.168.2.63, not on the Pi.
+# Set DSM proxy read/send timeout to 600s (default 60s returns HTTP 504).
 
 server {
     listen 80;
@@ -46,5 +48,6 @@ server {
         proxy_set_header   Upgrade $http_upgrade;
         proxy_set_header   Connection "upgrade";
         proxy_cache_bypass $http_upgrade;
+        proxy_read_timeout 600s;
     }
 }

nginx/kansho.conf

@@ -1,6 +1,8 @@
 # Kanshō – Host nginx (outside Compose)
 # Install as /etc/nginx/sites-available/kansho and symlink into sites-enabled.
 # TLS via nginx/certbot-setup.sh. Reverse-proxy targets are the Pi publish ports.
+# Live TLS currently terminates on the Synology Reverse Proxy at 192.168.2.63;
+# set DSM proxy read/send timeout to 600s (default 60s returns HTTP 504).
 
 server {
     listen 80;
@@ -50,5 +52,6 @@ server {
         proxy_set_header   Upgrade $http_upgrade;
         proxy_set_header   Connection "upgrade";
         proxy_cache_bypass $http_upgrade;
+        proxy_read_timeout 600s;
     }
 }